Out-of-distribution detection on CIFAR-10 (ID) vs SVHN (Far OOD) (test)
99.5AUROCGram matrix
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Gram matrixBackbone=Wide ResNet-502020.07 | 99.5 | — | |
| Contrastive Training for OOD DetectionBackbone=Wide ResNet-502020.07 | 99.5 | — | |
| Gram MatricesHyp.-Para. Dependency=No, Generalizable=true, Zero Shot=true2021.06 | 99.5 | — | |
| MahalanobisBackbone=Wide ResNet-50, Uses labeled OOD data for tuning=true2020.07 | 99.1 | — | |
| Residual flowsBackbone=Wide ResNet-50, Uses labeled OOD data for tuning=true2020.07 | 99.1 | — | |
| MahalanobisHyp.-Para. Dependency=High, Generalizable=false, Zero Shot=false2021.06 | 99.1 | — | |
| Rotation pred.Backbone=Wide ResNet-50, Uses additional data for pretraining=true2020.07 | 98.9 | — | |
| Outlier ExposureHyp.-Para. Dependency=Low, Generalizable=true, Zero Shot=false2021.06 | 98.76 | — | |
| Outlier exposureBackbone=Wide ResNet-50, Uses labeled OOD data for training=true2020.07 | 98.4 | — | |
| GeometricHyp.-Para. Dependency=Low, Generalizable=true, Zero Shot=true2021.06 | 97.96 | — | |
| CSI-ensHyp.-Para. Dependency=Low, Generalizable=true, Zero Shot=true2021.06 | 97.38 | — | |
| ODINBackbone=Wide ResNet-50, Uses labeled OOD data for tuning=true2020.07 | 96.7 | — | |
| ODINHyp.-Para. Dependency=High, Generalizable=false, Zero Shot=false2021.06 | 96.7 | — | |
| shifting transformation learningHyp.-Para. Dependency=No, Generalizable=true, Zero Shot=true2021.06 | 96.6 | — | |
| KNNk=102023.09 | 95.13 | 33.32 | |
| SSDHyp.-Para. Dependency=No, Generalizable=true, Zero Shot=true2021.06 | 93.8 | — | |
| Softmax probs.Backbone=Wide ResNet-502020.07 | 89.9 | — | |
| NNGuidealpha=10%, k=102023.09 | 89.62 | 43.64 | |
| Energy2023.09 | 88.39 | 44.94 |